The Value Of Multi-parameter MRI Radiomics In Predicting IDH-1 Mutation Status Of Glioma

Objective:To investigate the value of radiomics models established by preoperative conventional MRI sequences and diffusion kurtosis imaging sequences in predicting isocitrate dehydrogenase-1 gene mutation status in glioma.Methods:Clinical and imaging data were collected from 43 patients with glioma treated in our hospital from June 2014 to December 2019.Among them,15 cases were IDH-1 mutant-type and 28 cases were IDH-1 wild-type.Surgical sampling pathology was taken as a gold standard for IDH-1 gene mutation status.All patients were given preoperative MRI sequences scan,including T1WI,T2WI,T2-Flair,CE-T1WI and DKI sequences.Mean kurtosis(Mk),axial kurtosis(Ka)and fractional anisotropy(Fa)sequence were obtained after post-processing of DKI sequences.The obtained clinical and imaging data were uploaded to the Radcloud platform(Beijing Huiying Medical Technology Co,Ltd)for further analysis.Based on CE-T1WI and T2-Flair sequences,the tumor’s region of interest(ROI)were manually sketched layer by layer.After establishing the ROI,each sequence extracted 1409 imaging features.We labeled as IDH-1(negative/positive),and a five-fold cross-validation method was used for feature selection and model development.Then the variance threshold,SelectKBest method and Least Absolute Shrinkage and Selection Operator(LASSO)method were used for feature selection.A machine learning model constructed by support vector machine(SVM)algorithm was used for predicting IDH-1 mutation status.Individual sequence models and combined sequences models were evaluated separately,expressed as subject operating characteristic curves(ROC),with under curve(AUC),sensitivity,specificity and accuracy as evaluation indexes.The best prediction model was selected among all joint models.Results:Among the model for predicting IDH-1 mutation status,the mean AUC of test set of the machine learning model on T1WI to predict the mutation status of glioma IDH-1 was 0.61.The accuracy was 0.57,and the sensitivity and specificity were 0.74 and 0.27,respectively.The mean AUC of the test set of the machine learning model based on T2WI to predict the mutation status of glioma IDH-1 was 0.61.The accuracy was 0.65,and the sensitivity and specificity were 0.74 and 0.47,respectively.The mean AUC of the test set of the machine learning model based on T2-Flair to predict the mutation status of glioma IDH-1 was 0.68.The accuracy was 0.63,and the sensitivity and specificity were 0.69 and 0.53,respectively.The mean AUC of machine learning model based on CE-T1WI to predict the mutation status of glioma IDH-1 was 0.69.The accuracy was 0.71,and the sensitivity and specificity were 0.89 and 0.50,respectively.The mean AUC of the machine learning model based on Mk sequence for predicting the mutation status of glioma IDH-1 was 0.73.The accuracy was 0.77,and the sensitivity and specificity were 0.93 and 0.51,respectively.The mean AUC of the machine learning model based on Ka for predicting the mutation status of glioma IDH-1 was 0.69.The accuracy was 0.74,and the sensitivity and specificity were 0.82 and 0.60,respectively.The mean AUC of the machine learning model based on Fa to predict the mutation status of glioma IDH-1 was 0.66.The accuracy was 0.61,and the sensitivity and specificity were 0.65 and 0.53,respectively.The best combined model was constructed from Fa,Mk and T2-Flair feature,and the mean AUC of the combined model was 0.84,the accuracy was 0.81,and the sensitivity and specificity were 0.93 and 0.60,respectively.Conclusion:1.The radiomic machine learning model built by preoperative MRI sequences have clinical value in predicting glioma IDH-1 mutation status.2.The combination features of conventional MRI sequences and DKI sequences can improve the efficiency of the model.3.The feature extracted from DKI sequences have a better efficacy,and it preliminarily highlighted the potential of DKI as a non-invasive alternative method of predicting molecular in brain glioma evaluation.